System and method for deep learning recommender

This application claims priority from U.S. Provisional Patent Application No. 62/928,021 filed on Oct. 30, 2019, the entire contents of which are hereby incorporated by reference herein.

This disclosure relates to recommender systems, in particular, the application of deep learning techniques to generate recommendations.

Traditional recommender systems utilize information filtering to present information, such as a recommended product or service, that is likely to be of interest to a user. However, it can be challenging to generate useful recommendations in domains that have few products and services from which a recommendation can be drawn.

According to an aspect, there is provided a computer-implemented method for generating recommendations for users, comprising: identifying items held by the users, each of the items defined by a shallow representation and one or more attributes; defining the items held by the users based on a deep representation, the deep representation derived from attributes of the items; generating a deep holding matrix identifying the items held by the users based on the deep representation of the items; generating a shallow holding matrix identifying the items held by the users based on the shallow representation of the items; generating a similarity score matrix of scores between the deep representations based on a similarity between attributes or derived attributes; decomposing, by way of matrix factorization, the shallow holding matrix into a user latent representation and a product feature latent representation, based at least in part on the similarity score matrix; concatenating the product feature latent representation and product information to be passed to a first neural network to generate a first output; concatenating the user latent representation and user information to be passed to a second neural network to generate a second output; and performing a dot product matrix multiplication on the first output and the second output to generate, for every user and every product, a probability for a recommendation.

According to another aspect, there is provided a computer system comprising: a processor; a memory in communication with the processor, the memory storing instructions that, when executed by the processor cause the processor to perform a method as disclosed herein.

According to a further aspect, there is provided a non-transitory computer-readable medium having computer executable instructions stored thereon for execution by one or more computing devices, that when executed perform a method as disclosed herein.

According to yet another aspect, there is provided a computer-implemented method for generating a product recommendation for a user, comprising: particularize information about a plurality of products based on attributes of the products, one of the attributes being family; generate a holding matrix to identify user holdings of the products; calculate a similarity score matrix between the products based on common attributes between the products; generate a latent representation of the families of the products based on the holding matrix and the similarity score matrix; generate a latent representation of the user holdings of the products based on the holding matrix and the similarity score matrix; and using a neural network with input from information associated with each user, the product attributes, the latent representation of the families, and the latent representation of then users, generate the product recommendation for the user.

Other features will become apparent from the drawings in conjunction with the following description.

In embodiments disclosed herein, using deep learning techniques, a recommender system identifies potential products and/or services for a particular user or client. References to “products” and “services” as used herein will be understood to refer products and/or services, and apply interchangeably. “Products” and “services” may also be interchangeably referred to herein as “items”. Systems and methods disclosed herein may identify a client's information including their profile, size, age as well as financial health. Product profiling can also be performed. The recommender system can thus take into consideration the nature of products, user profile, and how other users have succeeded with their products to generate a recommendation to match a specific product to a client.

Embodiments of a recommender system disclosed herein may allow for recommendations to be made in domains, such as banking, that have fewer products and services available in certain domains.

Furthermore, embodiments of a recommender system disclosed herein may provide recommendations that account for more granular attributes of a product and/or service that is being recommended.

1 FIG.A 1 FIG.B 100 100 illustrates a recommender system, in an embodiment.illustrates an example data flow of recommender system, in an embodiment.

100 111 112 114 116 118 110 Recommender systemincludes a divider, a holding matrix constructor, a similarity score generator, a latent representation generator, a neural network, and a data store, as described in further detail below.

110 100 101 102 103 104 105 106 108 Data storecan be configured to store data used and generated by recommender systemsuch as a user profile, attributes, a holding matrix, a similarity score matrix, a product latent representation, a user latent representation, and a dataframe.

Products or services may be characterized by “shallow” products/services or “families”. In the banking industry, by way of example, various types of financial products such as mortgages, loans or lending, credit cards, and overdrafts, would all be examples of shallow products or families of products. A “shallow product” or “shallow representation” can be considered as a shallow categorization or representation of a product.

111 102 111 102 Dividermay be configured to divide shallow products/services into multiple “deep” products or services based on further particularized attributesof the products or services. Thus, dividermay be used to further particularize a product/service into more detailed granularity, namely, “deep products”, based on and identified by specific attributes of that product/service. A “deep product” or “deep representation” can be considered a derived attribute of a product, and derived from attributesof a product. It should be understood that a particular product and/or service can be defined both by its “shallow product” (e.g., loan) and/or as a “deep product” based on derived attributes (e.g., loan with a principal in a range of <$1000).

102 102 Division from shallow to deep products/services can be based in part on domain and business knowledge, for example, to identify the number of attributesand characteristics of those attributesto divide shallow products/services into.

102 Such division may be performed to differentiate products or services based on attributesthat are statistically significant, such as an interest rate in the context of a credit card.

102 In some embodiments, attributesinclude information related to deep product/service info, and the nature of such products or services.

102 In an example, a mortgage may be divided into separate products, each “deep product” thus differentiated by attributessuch as amount, amortization period, interest rate, and type of interest rate.

102 In some embodiments, distributions of attributesare determined using a suitable statistical technique such as quantile bucketing. To perform quantile bucketing, users holding a shallow product are sorted, for example, in an ascending or descending order based on values of a particular attribute of the shallow product (in an example, the shallow product is a loan and the attribute is the loan amount). The users are then distributed into a number of bins or buckets, such that each bucket contains the same (or almost the same) number of users. In order for each bucket to contain the same number of users, some buckets may cover a different span of values of the attribute (for example, a range of values for the loan amount).

The number of users that own a shallow product may be statistically divided into equal buckets, where each bucket will have almost similar number of users. The result is multiple buckets with different value ranges for the attribute.

Following a quantile bucketing procedure, the resulting attribute value ranges of each bucket can define a derived attribute, which can be used to form a “deep product” representation of a product.

In an example, a shallow product can be a loan, and a population of a hundred users who each hold a loan. The users can be sorted in ascending order of an attribute such as loan amount. The sorted users can be divided into five buckets, such that each bucket contains twenty users, and each bucket spans the range of loan amount values of the users in that bucket, such as <$10,000, $10,000-$25,000, and the like.

Such divisions may allow for an increase in the scope of number of products or services.

2 FIG.A 2 FIG.A 2 FIG.A 2 FIG.A 102 102 is an example of a table of values of attributesin the form of a deep products table.is a representation of how products can be divided into a “deeper” version (more granular) using features such as those shown (e.g., currency, interest type, limit and rate). Each row inrepresents a deep product. The “In_desc_grp” column lists “shallow products”, or a shallow representation of a product, such as “RBCCL” in an example. The first two rows ofillustrate shallow product RBCCL with different features where the LMT_B (the principal amount) is different (one between 0.09-5000 and 5000-1000. LMT_B is an example of a derived attribute, based on the attributeof each product. For example, the first row has a principal amount in the range of (5001-10000.0], where a square bracket can indicate that the end of range is inclusive, and a rounded parenthesis can indicate that the end of the range is exclusive. The column “Rate” indicates a derived attribute of the range of interest values that the interest rate for that particular deep product is in.

100 112 103 101 101 Recommender systemalso includes holding matrix constructor, which is configured to generate a holding matrixlisting each user's product holdings, based in part on their user profile. Each user is associated with a user profile, containing data such as client information, associated with users.

112 In some embodiments, holding matrix constructorgenerates a holding matrix for shallow products, containing information relating to which user's hold which shallow products.

112 In some embodiments, holding matrix constructorgenerates a holding matrix for deep products, containing information relating to which user's hold which deep products. A shallow holding matrix can be used for generating a latent representation, as described further below. A deep holding matrix can be used to make a model learn that a specific user likes a very specific product (deep product).

2 FIG.B illustrates an example of a deep holding matrix, in an embodiment.

114 104 114 102 A similarity score generatoris configured to calculate a similarity score matrixcontaining similarity scores between deep products. For example, similarity score generatorcan determine a percentage similarity between each deep product (based on attributes) using each deep product's attributes. In some embodiments, a similarity score may be calculated using domain knowledge provided by business partners, such as Business Financial Services (BFS).

In some embodiments, a similarity score between two deep products may be determined using a distance similarity scoring system, such as based on a binary distance measure between attributes.

In some embodiments, any deep product from different shallow product will have similarity scores of zero (indicating that there is not any similarity).

For deep products that are part of the same shallow product, a similarity score may be calculated based on how many attributes or features the deep products have in common, and defined as the proportion of attributes or derived attributes that the deep products share that have the same or similar values (within a threshold). In an example, if two deep products have seven attributes each, and one attribute value is different between two deep products, the similarity score can be: 6/7=85.7%, resulting in a similarity score of 85.7%. In an example of deep products with differences between two attribute values, the similarity score can be defined as 5/7=71.4%. In a further example, if two out of five attribute values are similar, two deep products may have a similarity score of ⅖=40%.

In some embodiments, other suitable similarity scoring functions can be used, including more complex similarity scoring such as cosine similarity based on attribute values between deep products.

In some embodiments, a similarity score may take into consideration, for different attribute values between deep products, the value or quantity of the difference between the attribute values, in combination or in place of how many attributes are different.

Similarity may be calculated between any product and all other products.

In some embodiments, a deep product compared to another deep product from a different family would have zero similarity, for example a credit card and a mortgage.

2 FIG.C illustrates an example of deep product similarity score, according to an embodiment.

116 102 105 105 In some embodiments, latent representation generatoris configured to map “deep product” data of attributesto a latent space of “shallow products”, generating a product latent feature representationsuch that “deep products” belonging to similar “shallow products” are closer in the latent space, or a shallow product representation. Product latent feature representationis thus a latent representation for a shallow product or family of products.

105 103 104 In some embodiments, product latent feature representationis generated based on holding matrixand similarity score matrix. In an example, dimensionality of “deep product” information may be reduced from one-hot encoding.

105 103 103 118 105 In some embodiments, product latent feature representationis generated by applying an embedding model such as matrix factorization on holding matrix, such as holding matrixfor shallow products. A shallow product holding matrix can be used for the latent representation and a deep product holding matrix can be used for inputs to the deep learning model (such as neural network). Product latent feature representationprovides a compressed numeric representation of a product/service given users' holdings.

103 105 106 Applying matrix factorization can involve decomposing holding matrixinto the product of two lower dimensionality rectangular matrices—namely, product latent feature representationand user latent representation.

In some embodiments, matrix factorization is a method to be able to predict every unknown entry in the given matrix. Original matrix will have very few known items and a lot of unknown items. The first step of matrix factorization is to break the original (holding) matrix into two much smaller matrices that when they are multiplied they fill every single entry (unknowns) in the original matrix.

116 Thus, latent representation generatormay be configured to fixate the latent feature representation of each of the divided deep products to the feature representation of the family (shallow product) of that product itself. This may ensure that a user with the same shallow product (e.g., mortgage) will have similar representations in latent space even if there product attribute values are different (e.g., 5-year mortgage as compared to 3-year mortgage).

Conveniently, dividing the products (to “deep products”) and then fixating their latent representations may allow for more granular products yet similar representation.

In some embodiments, users with similar product holdings will be given closer scores and it can be viewed as a closer distance in a given multi-dimensional space. Similarly, products that co-exist together will often have similar holdings. For example, if product A and B exist together a lot in the users, then they will have similar embeddings.

116 106 In some embodiments, latent representation generatoris also configured to generate a user latent representationor client representation.

For example, deep products may be grouped close to each other in latent space if users tend to hold those types of products together.

106 103 104 In some embodiments, user latent representationis generated based on holding matrixand similarity score matrix.

106 103 103 In some embodiments, user latent representationis generated by applying an embedding model such as matrix factorization on holding matrix, such as holding matrixfor shallow products.

106 Conveniently, generating a user latent representationmay identify similarities between users. For example, users with similar shallow product holdings will be given closer scores, which can represent a closer distance in a multi-dimensional space. If users are grouped together, this means that when that representation is given to the model the difference between the two users are not too different.

100 118 105 106 In some embodiments, recommender systemuses latent features/representations (such as for the product family itself instead of the divided deep products) as input to neural networkto generate a recommendation. This can allow for similarities to be identified between products (e.g., product latent feature representation) and users (e.g., user latent representation) with similar deep products.

118 The architecture of neural networkcan include a suitable number of networks or layers.

118 In some embodiments, neural networkincludes three connected networks or layers, in an embodiment, a first dense layer, a second dense layer and a third dense layer.

A first network or dense layer processes products/services by concatenating product/service features and embeddings.

A second network or dense layer processes the clients by concatenating user features and embeddings.

In some embodiments, the first network and the first network have identical structures.

A third network or dense layer processes the output of the first and second network.

118 In some embodiments, hyperparameters of neural networkcan be determined from testing using grid search where almost every combination was determined. Hyperparameters can include, for example, dropout rate, number of neurons in each layer, number of layers, batch_size, number of epochs, and the like. A grid search can involve trying almost every possible value/combination of the values of the hyperparameter to be able to find the one with the highest score.

101 106 102 105 118 In some embodiments, one or more of user profile, user latent representation, product attributes, product latent feature/representation, are input to neural network.

In some embodiments, there are three networks. One network takes the user information and user embeddings, another takes the items (products/services) information and item embeddings. The output of each of these networks are then fed to the third network that outputs a single score.

118 101 106 The first network of neural networkmay take user profile, user latent representationas input. In some embodiments, the activation function used in the first network is relu (“Recitified Linear Unit”) where the output is a 64-dimensional vector consisting of numbers that represents the users with all its information and latent representation.

118 102 105 The second network of neural networkmay take product attributesand product latent feature/representationas input. In some embodiments, the activation function used in the first network is ReLU (“Rectified Linear Unit”) where the output is a 64-dimensional vector consisting of numbers that represents the deep product with all its information and latent representation.

118 The third network of neural networkmay take the output of the first network and the second network as input to output a probability of a client holding that deep product. In some embodiments, the activation function is a sigmoid function resulting in a single number that is between 0 and 1 representing a matching score or probability of a given user and a specific deep product. In some embodiments, the output is probability/matching score for every user with every deep product. The score will be the probability that a user will like a product/service.

118 Neural networkcan be trained using a suitable training technique.

Following training, a traditional neural network prediction or inference function can take minutes to execute, affecting performance. In some embodiments, output trained embeddings are obtained from the output of the first (user) neural network and the second (product) neural network. and instead of being passed to the third neural network, a dot product operation is performed, performing matrix multiplication between the output of the first and the second neural network, generating every combination between users and products/services in a matter of seconds.

A trained model prediction can be served or provided, for example, to a user, by looking up the user-item score. The user-item score can indicate a probability that a user will like a product/service. All the predictions can be saved and given to the user.

2 FIG.D 118 illustrates an architecture for neural network, in an embodiment. Product information and product embeddings are concatenated to form a product train that is passed to the first network. The first network may have three layers, of 64 nodes using ReLU activation functions. Client information and client latents are concatenated to form a client train that is passed to the second network. The second network may have three layers, of 64 nodes using ReLU activation functions. A dot product operation is performed on the output of the first network and the second network, generating the final output.

2 FIG.E illustrates an example of the final output, in an embodiment.

1 FIG.B As illustrated in, an output of the model will be for each given client/user, a top-n recommendation of deep products where n is a pre-defined number, for example, chosen by the business, along with the probability of the likelihood that the given user will like the recommended product.

108 108 Thus, a dataframecan be constructed for each user with n number of products that are similar to a user's existing holdings. The score for each product, for example, between 0-100, is the maximum score between a randomly chosen product and the product held by the user. Similarities of zero may be omitted from dataframe.

100 In some embodiments, recommender systemdoes not find the closest product or service for a user. Instead, n number of products are chosen using a random sampling technique. In an example, this can be negative sampling. Random products that do not belong to the users are chosen from all the products that are not owned by the user to be fed to the model. Conveniently, this may save a lot of time and computation power.

In some embodiments, a probability of client holding is represented by deep product similarity score.

3 FIG.A 300 illustrates a flow chart of a methodfor deep learning based recommendation, according to an embodiment. The steps are provided for illustrative purposes. Variations of the steps, omission or substitution of various steps, or additional steps may be considered.

302 111 102 At block, for shallow product or service data, dividerdivides shallow product information to deep product information, stored as attributes.

304 112 103 At block, holding matrix constructorgenerates holding matrix, identifying which users hold which products.

306 114 104 102 At block, similarity score generatorgenerates a similarity score matrixwith the percentage similarity between each deep product (based on attributes) using each product's attributes.

308 116 105 103 104 At block, latent representation generatorgenerates product latent feature/representationbased on holding matrixand similarity score matrix, a latent representation for the shallow product/family of products.

310 116 106 103 104 At block, latent representation generatorgenerates user latent representationis generated based on holding matrixand similarity score matrixto identify similarities between users.

312 118 101 106 102 105 At block, neural networktakes user profile, user latent representation, product attributes, product latent feature/representationto generate a top-n recommendation of deep products for a user.

It should be understood that one or more of the blocks may be performed in a different sequence or in an interleaved or iterative manner.

3 FIG.B 320 illustrates a flow chart of a methodfor deep learning based recommendation, according to an embodiment. The steps are provided for illustrative purposes. Variations of the steps, omission or substitution of various steps, or additional steps may be considered.

321 At block, items held by the users are identified, each of the items defined by a shallow representation and one or more attributes.

322 At block, the items held by the users are defined based on a deep representation, the deep representation derived from attributes of the items.

323 At block, a deep holding matrix is generated, identifying the items held by the users based on the deep representation of the items.

324 At block, a shallow holding matrix is generated, identifying the items held by the users based on the shallow representation of the items.

325 At block, a similarity score matrix of scores is generated between the deep representations based on a similarity between attributes or derived attributes.

326 At block, the shallow holding matrix is decomposed, by way of matrix factorization, into a user latent representation and a product feature latent representation, based at least in part on the similarity score matrix.

327 At block, the product feature latent representation and product information are concatenated and passed to a first neural network to generate a first output.

328 At block, the user latent representation and user information are concatenated and passed to a second neural network to generate a second output.

329 At block, a dot product matrix multiplication is performed on the first output and the second output to generate, for every user and every product, a probability for a recommendation.

It should be understood that one or more of the blocks may be performed in a different sequence or in an interleaved or iterative manner.

100 111 112 114 116 118 110 120 300 302 312 320 321 329 120 4 FIG. Recommender system, in particular, one or more of divider, holding matrix constructor, similarity score generator, latent representation generator, neural network, and data store, may be implemented as software and/or hardware, for example, in a computing deviceas illustrated in. Method, in particular, one or more of blocksto, and method, in particular, one or more of blocksto, may be performed by software and/or hardware of a computing device such as computing device.

120 210 220 230 240 250 As illustrated, computing deviceincludes one or more processor(s), memory, a network controller, and one or more I/O interfacesin communication over bus.

210 Processor(s)may be one or more Intel x86, Intel x64, AMD x86-64, PowerPC, ARM processors or the like.

220 Memorymay include random-access memory, read-only memory, or persistent storage such as a hard disk, a solid-state drive or the like. Read-only memory or persistent storage is a computer-readable medium. A computer-readable medium may be organized using a file system, controlled and administered by an operating system governing overall operation of the computing device.

230 Network controllerserves as a communication device to interconnect the computing device with one or more computer networks such as, for example, a local area network (LAN) or the Internet.

240 120 230 One or more I/O interfacesmay serve to interconnect the computing device with peripheral devices, such as for example, keyboards, mice, video displays, and the like. Such peripheral devices may include a display of device. Optionally, network controllermay be accessed via the one or more I/O interfaces.

210 220 240 210 210 Software instructions are executed by processor(s)from a computer-readable medium. For example, software may be loaded into random-access memory from persistent storage of memoryor from one or more devices via I/O interfacesfor execution by one or more processors. As another example, software may be loaded and executed by one or more processorsdirectly from read-only memory.

220 120 120 Example software components and data stored within memoryof computing devicemay include software to generating a deep learning based recommendation, as described herein, and operating system (OS) software allowing for basic communication and application operations related to computing device.

Embodiments as disclosed herein may advantageously improve accuracy in identifying potential needs for clients. For example, accuracy can be measured using AUC (Area Under the Curve), RMSE (Root Mean Squared Error). Earlier models' AUC was averaged to be 82.7% and with the deep learning model score AUC of more than 96% which means that the model has improved a lot over earlier techniques in identifying potential needs for clients.

Conveniently, embodiments as disclosed herein may provide detailed information of a product (deep product), can save three to four days of paper work by the account manager, and improve accuracy of product predictions by at least 13%.

Of course, the above described embodiments are intended to be illustrative only and in no way limiting. The described embodiments are susceptible to many modifications of form, arrangement of parts, details and order of operation. The disclosure is intended to encompass all such modification within its scope, as defined by the claims.